Geodesic trajectory generation on learnt skill manifolds
Humanoid robots are appealing due to their inherent dexterity. However, their potential benefits may only be realized with a correspondingly flexible motion synthesis procedure. Designing flexible skill representations that also capture non-trivial dynamics effects over a large domain, such as in real humanoid robots, has been an open challenge. This poster presents one such flexible trajectory generation algorithm that utilizes a geometrical representation of humanoid skills (e.g., walking) - in the form of skill manifolds. Such manifolds are learnt from demonstration data that may be obtained from off-line optimization algorithms (or a human expert). This model may be used to produce approximately optimal motion plans (that capture constraints and dynamics implicit in the output of a computationally expensive off-line optimization procedure) as geodesics over a manifold and this allows us to effectively generalize from a limited training set. We demonstrate the effectiveness of our approach on a physical 19-DoF humanoid robot, exhibiting fast motion planning on a realistic – variable step length, width and height – walking task.